[ros-users] Reinventing the wheel: training and evaluation of algorithms for objects detection in images
Gary Bradski
bradski at willowgarage.com
Mon Feb 28 22:34:47 UTC 2011
I had missed this in my email, don't know where you are yet. eblearn looks
cool but all these things have limits that work well on certain classes of
problems and not others (that is in fact a consequence of the "no free lunch
theorem"). I sort of like FLANN because it is easy to scale over clusters,
easy to load local models, is inherently incremental and on a give machine
it scales logarithmically in the number of objects learned. The user's job
would then be not to find "good" features but features that work well with
FLANN.
For the solutions in perception challenge. I had thought about not requiring
people to use ROS. My next thought was: "Let's require people to use ROS or
else they are free not to join". The reason is, someone else is providing
and maintaining a rich framework ROS and it's uniform so that someone can
compare their new algorithms against old ones without changing API etc.
It's also nice in that once you write code for the contest, the exact same
code can be used to grab objects using a real robot or the robot simulator.
A lot of things like that come for free on ROS including dynamic settings of
parameters etc.
One difference that you might think of is that I don't want to just identify
an object, I want to know its pose relative to the camera.
Gary
On Tue, Nov 23, 2010 at 3:06 PM, Pierre S <pierre.sermanet at gmail.com> wrote:
>
>
>> All and all eblearn + visiongrader seem like great tools, with a nice
>> and well thought API.
>> I will definitely look more in detail about their internals (I already
>> had played around with eblearn a few months ago) and use/re-use as
>> much as possible the existing code base.
>>
>>
> Thanks for your interest Rodrigo. If you reuse existing parts, I guess the
> best for what you want to do would be to use the tensor library libidx and
> its GUI as a basis for image/matrix manipulations (not tied to any learning
> scheme) and visiongrader for performance evaluation.
>
> Pierre
>
> _______________________________________________
> ros-users mailing list
> ros-users at code.ros.org
> https://code.ros.org/mailman/listinfo/ros-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osuosl.org/pipermail/ros-users/attachments/20110228/95fcd154/attachment-0002.html>
More information about the ros-users
mailing list